Derivation and validation of a machine learning risk score using biomarker and electronic patient data to predict progression of diabetic kidney disease

• DOI: 10.1007/s00125-021-05444-0
• Publication Date: 2021-04-02T11:02:52Z
• Authors: Lili Chan; Girish N. Nadkarni; Fergus Fleming; James R. McCullough; Patricia Connolly; Gohar Mosoyan; Fadi El Salem; Michael W. Kattan; Joseph A. Vassalotti; Barbara Murphy; Michael J. Donovan; Steven G. Coca; Scott M. Damrauer
• Keywords: Adult; Male; Kidney Function Tests; Article; Cohort Studies; Machine Learning; Young Adult; 03 medical and health sciences; 0302 clinical medicine; Predictive Value of Tests; Risk Factors; Electronic Health Records; Humans; Diabetic Nephropathies; Aged; Aged, 80 and over; Middle Aged; Prognosis; United States; 3. Good health; Disease Progression; Female; Biomarkers; Glomerular Filtration Rate

Abstract Aim Predicting progression in diabetic kidney disease (DKD) is critical to improving outcomes. We sought develop/validate a machine-learned, prognostic risk score (KidneyIntelX™) combining electronic health records (EHR) and biomarkers. Methods This an observational cohort study of patients with prevalent DKD/banked plasma from two EHR-linked biobanks. A random forest model was trained, performance (AUC, positive negative predictive values [PPV/NPV], net reclassification index [NRI]) compared that clinical Kidney Disease: Improving Global Outcomes (KDIGO) categories for predicting composite outcome eGFR decline ≥5 ml/min per year, ≥40% sustained decline, or failure within 5 years. Results In 1146 patients, the median age 63 years, 51% were female, baseline 54 ml min −1 [1.73 m] −2 , urine albumin creatinine ratio (uACR) 6.9 mg/mmol, follow-up 4.3 years 21% had endpoint. On cross-validation derivation ( n = 686), KidneyIntelX AUC 0.77 (95% CI 0.74, 0.79). validation 460), 0.76, By comparison, 0.62 0.61, 0.63) 0.61 0.60, validation. Using cut-offs, stratified 46%, 37% 17% into low-, intermediate- high-risk groups endpoint, respectively. The PPV progressive function group 61% vs 40% highest strata by KDIGO categorisation p < 0.001). Only 10% those scored as low experienced (i.e., NPV 90%). NRI event 41% 0.05). Conclusions improved prediction outcomes over models individuals early stages DKD. Graphical abstract